Researches On Spectral-spatial Feature Extraction And Classification Methods For Hyperspectral Remote Sensing Imagery

Remote sensed hyperspectral images can provide rich and reliable spectral information.Thus, hypersectral remote sensing plays an important role in earth observation system and iswidely used in military, accurate agriculture, environmental monitoring, and other related fields.However, when dealing with hyperspectral images, the traditional image processing methodsare confronted with new challenges, such as classification with limited training samples. Thisdissertation deeply analyzes the characteristics of hyperspectral images and systematically re-views current hyperspectral processing works. Multiple novel hyperspectral image feature ex-traction and classification technologies are proposed in this dissertation. The efectiveness of theproposed methods are demonstrated in multiple real hyperspectral remote sensing applications.For hyperspectral image feature extraction, this paper proposes a method based on imagefusion and recursive filtering. In the proposed feature extraction method, image fusion and re-cursive filtering are used to remove image noise and enhance the salient spatial structures ofimages, such as shapes and contours. This method can increase dissimilarities of the pixels be-longing to diferent objects while decreasing dissimilarities of the pixels belonging to the sameobject. Experimental results demonstrate that the features obtained by the proposed method canefectively represent the spatial contextual information of objects, and thus, can significantlyimprove the accuracies of pixel-wise classifiers such as support vector machines.Spectral reflectance is the key feature for analyzing hyperspectral images. However, thespectral reflectance recorded by remote sensed hyperspectral images are usually afected bymany factors such as sensing mechanism, capturing circumstance, and weather conditions. Toreduce the interferences of these factors, this paper introduces a novel hyperspectral image fea-ture extraction method based on intrinsic image decomposition. This method is based on theperceptive function of human vision to distinguish illumination and spectral reflectance. Specif-ically, the hyperspectral image is first separated into two components, i.e., the illumination andreflectance components. Then, only the spectral reflectance component is used for the followingpixel-wise classification. The proposed method is compared with multiple recently proposedhyperspectral image feature extraction and classification methods. Experimental results showthat intrinsic image decomposition can greatly reduce the interferences of many factors suchas scene illumination, shading, and noise, and thus, the classification accuracy can be greatlyimproved especially when the number of training samples is quite limited.For the fast spectral-spatial classification of hyperspectral images, an edge-preserving fil-tering based hyperspectral image classification method is proposed which consists of the follow- ing steps: First, the pixel-wise classification result is obtained by the support vector machinesclassifier. Then, the pixel-wise classification map is represented as multiple probability mapsand optimized locally with edge-preserving filtering. The final classification map is obtained bymaximizing the probabilities. Edge-preserving filter is a type of non-linear classifier. Comparedwith the Gaussian filter, the edge-preserving filter can further consider the spatial information ofhyperspectral images in the filtering process. Two non-iterative edge-preserving schemes, i.e.,bilateral filter (extended from Gaussian) and guided filter (based on local linear model) are usedin this work. The AVIRIS and ROSIS-3data sets are used in the experiments to compare theproposed method with recently proposed spectral-spatial classification methods. Experimentalresults show multiple advantages of the proposed method, such as low computing complexity,high accuracy and so on.Local optimization based spectral-spatial classfication methods are not able to model thedeep spatial contextual correlation in hyperspectral images, and thus, not able to obtain a highclassification accuracy when the number of training samples is quite limited. In order to solvethis problem, a novel spectral-spatial classification method based on extended random walkersis proposed. Through extending the original random walkers, an energy function which fuse thespatial and spectral information of hyperspectral images is constructed. The energy function hasclosed form solutions, and thus, can be easily solved to obtain the classification result. Exper-imental results show that the proposed method can achieve breakthrough on recently proposedspectral-spatial methods when the number of training samples is quite limited.Finally, four real case studies are used to test the multiple feature extraction and classifi-cation methods proposed in this dissertation：1) Based on the proposed extend random walkersbased spectral-spatial classification method, the long-wave infrared hypersectral images andvisible images are fused to classify the land objects in Black Lake area of Thetford Mines,province of Que′bec, Canada;2) The proposed feature extraction and classification methods areused to classify the vegetation around the Kennedy Space Center and plot the spatial distributionmap of the vegetation in this area;3) The proposed methods are used to classify the vegetationin Botswana wetland and make detection maps of the water and fire-scare in this area;4) Theproposed methods are applied to classify and plot a spatial distribution map of water, soil, andvegetation around the Yellow River Delta, China.